pipeline-content

Content creation pipeline — research, strategy, create, critique, humanize, knowledge capture. Chains auto-researcher, content-strategy, content-creator/copywriting/tech-writer, content-critique, huma

SKILL.md

GitHub

Pipeline: Content Creation

You are orchestrating a content creation pipeline. Every piece of content goes through research, creation, critique, and humanization before delivery.

Input Parameters

Collect these before starting (use AskUserQuestion for any missing):

  • topic: What the content is about
  • type: blog | social | email | docs | landing | ad | video-script
  • platform (for social): tiktok | instagram | facebook | linkedin | youtube | x | threads
  • audience: Who is this for?
  • brand/slug (optional): Brand identity to use
  • language (optional): en | vi (default: en)

Pipeline State

Create a tracker at .gstack/pipeline-content-{date}.md.

## Pipeline: Content Creation
Started: {timestamp}
Topic: {topic}
Type: {type}
Platform: {platform}

| # | Stage | Status | Gate | Notes |
|---|-------|--------|------|-------|
| 1 | RESEARCH | pending | — | — |
| 2 | STRATEGY | pending | — | — |
| 3 | CREATE | pending | — | — |
| 4 | CRITIQUE | pending | — | — |
| 5 | HUMANIZE | pending | — | — |
| 6 | KNOWLEDGE | pending | — | — |

Stage 1: RESEARCH

Invoke /auto-researcher with the topic.

The researcher will search, synthesize, and produce a structured report with:

  • Key findings with confidence levels
  • Supporting evidence with source quality tiers
  • Conflicting views
  • Limitations

Save the research output — it feeds into Stages 2 and 3.

Gate: Research report must have at least 3 data points with High or Medium confidence. If all sources are Low confidence, warn the user and ask whether to proceed.

On pass: Update tracker → PASS, proceed.


Stage 2: STRATEGY (conditional)

Skip if: type is social, ad, or email AND user didn't request a campaign.

Run if: type is blog, landing, docs, or user explicitly requests strategy.

Invoke /content-strategy with:

  • The research findings from Stage 1
  • The target audience
  • The content type and platform

Output: content brief with angle, outline, keyword targets, CTA, and distribution plan.

Gate: User reviews and approves the brief. If skipped, the research report serves as the brief.


Stage 3: CREATE

Select the creation skill based on content type:

Type Skill Why
social, ad, email, landing /content-creator Conversion-focused with formulas, psychology effects, platform constraints
blog, video-script /copywriting Long-form persuasive writing with AIDA/PAS frameworks
docs /tech-writer Developer documentation with Diataxis framework

Pass to the selected skill:

  • The research findings from Stage 1
  • The content brief from Stage 2 (if available)
  • All input parameters (audience, platform, language, brand)

Gate: Draft content produced. Present to user for initial review before critique.


Stage 3.5: COMPLIANCE (fintech only)

Trigger condition: Run this stage ONLY when ANY of these are true:

  • preset=banking-finance or preset=crypto is active
  • Content type involves financial products or services
  • Product type is neobank, payments, lending, insurance, wealthtech, or crypto
  • Target audience is in a regulated financial jurisdiction

Skip if: Content is general business/tech content with no financial product claims.

Invoke /fintech-compliance-gate with:

  • The draft content from Stage 3
  • Product type (infer from preset or ask user)
  • Target jurisdiction(s)
  • Distribution platform

The compliance gate runs a 20-item checklist and produces RED/AMBER/GREEN findings.

Gate logic:

  • All RED items resolved → must return to Stage 3 CREATE with compliance requirements attached
  • AMBER items → add required disclosures to the draft, then proceed
  • All GREEN → proceed to Stage 4

Update tracker:

| 3.5 | COMPLIANCE | fintech-compliance-gate | {PASS/FAIL/PASS_WITH_DISCLOSURES} | {RED count}/{AMBER count}/{GREEN count} | {notes} |

Stage 4: CRITIQUE

Invoke /content-critique on the draft.

The critique skill:

  • Grades A-F across 6 dimensions (Clarity, Accuracy, Tone, Structure, SEO/Value, Consistency)
  • Runs /stop-slop internally for AI pattern detection
  • Produces severity-tiered findings with exact locations

Gate: Grade B or above → proceed to Stage 5.

If grade C or below:

  1. Report the critique findings to the user
  2. Apply the Top 3 fixes recommended by the critique
  3. Re-run the creation skill with the fixes applied
  4. Re-run critique on the revised draft
  5. If still C after 2 revision cycles → present to user with the critique report and ask whether to proceed

Stage 5: POLISH

Invoke /content-polish on the critique-approved draft.

The content-polish skill orchestrates three passes in sequence (language-branched: English → humanizer/proofreader, Vietnamese → humanizer-vi/grammar-checker-vi with optional translationese-cleaner-vi pre-pass):

  1. Humanizer (format-calibrated) — removes AI patterns, calibrated to the document type
  2. Anti-fragmentation pass — catches over-fragmentation, restores connective tissue, preserves parallelism
  3. Proofreader / grammar check (post-humanizer mode) — catches typos, grammar, lost specifics, broken flow

This stage is MANDATORY per CLAUDE.md rules. Never skip.

Gate: Content-polish delivers a final "Version C" that passes:

  • Humanizer's two-pass self-audit (30 AI-pattern categories)
  • Anti-fragmentation check (no 3-short-sentences-in-a-row)
  • Proofreader's format-calibrated review (spelling, grammar, clarity, flow)

Stage 6: KNOWLEDGE (parallel, fire-and-forget)

Run these as background tasks after delivering the final content to the user:

6a: Knowledge graph

Invoke /graphify with the final content to capture entities, relationships, and topics in the knowledge graph.

This is fire-and-forget — don't block the pipeline report on it.


Final Report

## Pipeline Report: Content Creation
Topic: {topic}
Type: {type} | Platform: {platform}
Audience: {audience}
Run: {timestamp}

| # | Stage | Skill | Result | Gate | Notes |
|---|-------|-------|--------|------|-------|
| 1 | RESEARCH | auto-researcher | {result} | {N} data points | {sources used} |
| 2 | STRATEGY | content-strategy | {result/SKIPPED} | {brief approved} | — |
| 3 | CREATE | {skill used} | {result} | Draft produced | {angle/formula used} |
| 4 | CRITIQUE | content-critique | {result} | Grade: {letter} | {revision cycles} |
| 5 | HUMANIZE | humanizer | {result} | Self-audit: PASS | — |
| 6 | KNOWLEDGE | graphify | {result} | — | background |

Overall: {PASS / PASS_WITH_REVISIONS}
Final content: {delivered inline or file path}

Vietnamese Content Pipeline

When target is Vietnamese (language=vi): STRATEGY may load factual files from skills/vietnamese-language/ (seo-content-marketing.md, viral-content.md, platform files). CREATE does NOT load vietnamese-language — voice comes from exemplars (brand/voice-exemplars-vi/, pending), and the brief and outline stay in Vietnamese; load gen-z-slang.md only to check a term's meaning, never as a phrase source. CRITIQUE checks register consistency and AI-tells; HUMANIZE applies humanizer-vi patterns.